Defining Characteristics of QFabric
DEFININg CHARACTERISTICS OF QFABRIC
By Pradeep Sindhu
The purpose of this note is to introduce a set of defining characteristics
that set QFabric apart from other technologies that have been used or
proposed for building the internal networks of large scale data centers. It
is important to state that these characteristics complement each other
and work synergistically to help QFabric deliver its benefits. Thus we will
insist that for any technology to be labeled “QFabric” it must have all of the
defining characteristics and not just some of them.
The note begins with a set of definitions and a description of the general
benefits of QFabric. It then lists the defining characteristics and says why
they are important to building efficient, cost effective, dynamic, and easy
to manage data centers.
Definitions and General Benefits
QFabric is the name of a packet-switched networking technology
purpose-built to enable the construction of highly efficient, cost A QFabric has the fundamental property of scalability which is
effective, dynamic, and easy to manage data centers over a wide defined as the ability to maintain a set of defining characteristics
range of scales using standard off-the-shelf computing, storage, as the implementation grows in size from small values of N to
and services elements. The elements connect to QFabric via open large values of N. These defining characteristics are:
standard network interfaces such as Ethernet and Fiber Channel. 1. Any-to-any connectivity with fairness and full non-blocking:
QFabric will enable the performance of data centers to be improved This is the ability of a set of interfaces to send and receive
at a much faster rate than is possible by relying on performance packets to and from any other set with no restrictions or pre-
improvements of the infrastructure elements alone (e.g., Moore’s Law planning. Specifically, this includes the ability to absorb rapid
for microprocessors). We call this enhanced ability of QFabric to improve changes to the rate of transmission, the number of active
the performance of data centers over time “exponential scaling”. senders, and the number of active receivers. In all cases the
Overall, QFabric embodies two quintessential capabilities: full bandwidth of a target interface(s) is shared equally by
all contending interfaces instantaneously and continuously,
• The ability to treat data center computing, storage, services, and including in the special case of one interface sending to
network resources as fully fungible pools that can be dynamically just one other interface; this equal sharing is referred to as
and rapidly partitioned without the infrastructure or the applications “fairness”. Finally, the only apparent congestion is due to
knowing details about each other1. This is the key to simplicity, the limited bandwidth of ingress and egress interfaces and
efficiency, and security in the data center. any congestion of egress interfaces does not affect ingress
• The ability to connect the resources to each other at very high interfaces sending to non-congested interfaces; this non-
speeds with apparently no limitations in the interconnect except interference is referred to as “non-blocking”.
fixed interface bandwidth and a small transit latency. This is the key 2. Low latency and jitter: QFabric has interface-to-interface
to high performance as well as further efficiency improvements. latency on the order of 2 microseconds at small scale growing
A QFabric is represented as a circular object with N identical slowly to about 10 microseconds at the largest scale. The
interfaces to which server, storage, services, or network devices can latency also grows slowly with offered traffic load. QFabric also
attach themselves. The circular symbol is chosen to suggest the provides very low jitter.
property that all interfaces of QFabric are equidistant to each other in
3. No packet drops under congestion: When the instantaneous
terms of latency and equivalent in terms of bandwidth they can sink or
rate of packets coming in exceeds the instantaneous rate
source. This means that the internal network of QFabric appears to its
at which they are going out, QFabric signals the source
users to be flat, not hierarchical. The number of interfaces N is called
(e.g. servers or VMs) causing congestion to slow down by
the scale of the QFabric.
the appropriate amount so that the arrival rate matches
the departure rate. This throttling happens rapidly and
continuously to match input rate to output rate in a smooth •
Industry Standard Interfaces manner.
4. Linear cost and power scaling: The cost and power
consumption of QFabric increases linearly with the number of
interfaces N. This is in sharp contrast to traditional approaches
where both cost and power increase much faster than linearly.
5. Support of virtual networks and services: QFabric implements
virtual Layer 2 and Layer 3 networks to support multiple
tenants each running multi-tier applications. Complex security
and services requirements are supported by the insertion of
Layer 4-7 processing at any point in an application’s workflow.
Full mobility of virtual machines from any interface to any
Figure 1: A QFabric with 16 Interfaces
other interface is supported. Support of virtual networks does
not compromise any of the other properties.
Comparison to the Ideal Fabric 6. Modular distributed implementation that is highly reliable
and scalable: QFabric is built using a set of modular
We define the Ideal Fabric as one that has fixed interface bandwidth,
hardware and software components that are distributed and
but infinite internal bandwidth and zero interface-to-interface latency.
federated to provide high levels of redundancy. The modular
As we will see from the following defining characteristics, QFabric
implementation is designed to permit increasing or decreasing
comes very close to the Ideal Fabric. In other words, QFabric is about as
the number of interfaces while the system is running, a property
well as one can do given the constraints of physical implementation.
we call “dynamic scalability”.
7. Single logical device: Despite its distributed implementation,
QFabric acts as a single logical packet switching device—the
complexity of its distributed implementation is hidden without
removing any of the desirable properties such as high reliability
or dynamic scalability.
In a multi-tenant data center it is completely impractical to assume that the infrastructure knows
about the details of any application or that applications know about details of the infrastructure.
Why these characteristics are important
These defining characteristics represent a quantum jump in the networking technology used to build data centers.
They are the principal way in which we will provide exponential scaling for data centers for the next decade.
Importance of Scalability Importance of No Drop under Congestion
Scalability is the fundamental characteristic of QFabric. It is The ability to not drop packets when congestion occurs is critical
important for two reasons: to efficiently transporting “bursty” server to disk traffic. Most
applications assume that reads and writes to disk will succeed
• Economics: Scale is a pre-requisite for achieving better
with very high probability. Packet drops due to congestion breaks
economics in data centers. Small scale data centers simply
this assumption and forces the application to handle packet loss
cannot be made as efficient as large scale ones because of the
as an error, resulting in a drastic reduction in performance given
pooling principle, which states that when P equal size partitions
the relative frequency of congestion events. As noted earlier, in
of a resource are combined into a single pool, we need fewer
QFabric the no drop property is provided by not harming “innocent
units of the resource to provide the same level of service. Of
bystanders” interfaces whose traffic is not destined to congested
course, to reap the benefits of pooling we need to be able to
interfaces. This latter property of the implementation is key to
make the resources fully fungible.
supporting multi-tenancy and multiple applications within a tenant.
• Performance: For most applications the raw performance of a
set of tightly coupled computing elements in a single large data Importance of Linear Cost and Power Scaling
center is significantly better than the collective performance Since each interface requires some forwarding hardware, it is easy
of these same elements distributed over a number of smaller to see that the cost and power of a fabric must increase at least
data centers. This performance difference has to do with the linearly with the number of interfaces N. QFabric is able to achieve
inherently lower latency and higher bandwidth of inter-processor the ideal of linear scaling, which is highly desirable from both a
communication in a single data center. CAPEX and an OPEX standpoint.
Importance of Any-to-Any Fair, Non-Blocking Importance of Support for Virtual Networks and Services
Connectivity Support for Virtual Networks is important because data centers
This type of connectivity is critical to pooling the computing and are increasingly multi-tenant and have complex requirements for
storage resources in a data center. Without these capabilities application security, performance, and reliability. Virtual networks
resources in a large data center would remain stranded and not provide the basic tools to allow resources to be partitioned and
efficiently usable. This “flat” connectivity also goes a long way yet allow them to communicate securely when necessary. These
towards simplifying the writing of applications since there is no abstractions are essential to decoupling the applications from the
need to worry about the performance hierarchy of communication infrastructure.
paths inside the data center. It also relieves the operations staff
Importance of a Distributed, Modular Implementation
from having to worry about the “affinity” of application components
in order to provide good performance. A modular implementation where the modules are kept
independent through the use of physical or logical separation
Importance of Low Latency has two important benefits. The first is that failures in either the
Low latency in QFabric is critical to high performance, especially hardware or software are unable to compromise the entire system.
for modern applications where the ratio of communication to The second is that it becomes possible to increase or decrease the
computation is relatively high compared to legacy applications. For size of the system while it is running.
example financial applications are especially sensitive to latency.
Importance of Acting as a Single Logical Device
High latency translates directly to lower performance because
applications stall when they are waiting for a response over the This characteristic is important because it provides the simplest
interconnect—the more frequent and longer the stall, the lower the possible model for the network from a data center administrator’s
performance. standpoint. It permits orchestration and management applications
to be as simple as possible.
QFabric is a revolutionary new network technology purpose-built to allow the construction
of large scale data centers using off-the-shelf computing and storage elements connected
using industry standard network interfaces. It allows the performance of data centers to
be “scaled exponentially”, or improved at a rate much faster than is possible by relying on
improvements of the computing and storage elements alone. This exponential scaling
property is fundamental to making data centers not only much more powerful but also more
economically efficient, dynamic, and easier to manage.
At the highest level, QFabric embodies two quintessential capabilities. The first is its ability to
treat data center computing, storage, services, and network resources as fully fungible pools
that can be dynamically and rapidly partitioned without the infrastructure or the applications
knowing details about each other; this is the key to simplicity, efficiency, and security in data
centers. The second is the ability to connect the resources to each other at very high speeds
and with very low latency while ensuring fairness; this is the key to high performance and
further efficiency improvements.
This paper has laid out seven defining characteristics of QFabric that set it apart from all
other networking technologies currently in use or proposed for the data center. These defining
characteristics are orthogonal to each other and encompass performance, cost, power,
virtualization, security, reliability, and modularity dimensions that work together to provide the
two quintessential capabilities mentioned above.
QFabric represent a quantum jump in the networking technology used to build data centers
and will be the key to exponentially scaling high performance computing and storage over the
Vice Chairman, Chief Technical Officer and Founder
Pradeep Sindhu founded Juniper Networks in February 1996, and
is now the company’s Vice Chairman and CTO. His founding vision
was and still is to Connect Everything and Empower Everyone.
Before founding Juniper Networks, Dr. Sindhu was a Principal
Scientist and Distinguished Engineer at the Computer Science
Lab at Xerox’s Palo Alto Research Center (PARC).
He holds a bachelor’s degree in electrical engineering from the Indian Institute of Technology
in Kanpur, as well as a master’s degree in the same discipline from the University of Hawaii. In
addition, Dr. Sindhu holds both a master’s and a doctorate degree in computer science from
Carnegie Mellon University.
Corporate and Sales Headquarters APAC Headquarters EMEA Headquarters
Juniper Networks, Inc. Juniper Networks (Hong Kong) Juniper Networks Ireland
1194 North Mathilda Avenue 26/F, Cityplaza One Airside Business Park
Sunnyvale, CA 94089 USA 1111 King’s Road Swords, County Dublin, Ireland
Phone: 888.JUNIPER (888.586.4737) Taikoo Shing, Hong Kong Phone: 35.31.8903.600
or 408.745.2000 Phone: 852.2332.3636 EMEA Sales: 00800.4586.4737
Fax: 408.745.2100 Fax: 852.2574.7803 Fax: 35.31.8903.601
Copyright 2011 Juniper Networks, Inc. All rights reserved. Juniper Networks, the Juniper Networks logo, Junos,
NetScreen, and ScreenOS are registered trademarks of Juniper Networks, Inc. in the United States and other
countries. All other trademarks, service marks, registered marks, or registered service marks are the property of
their respective owners. Juniper Networks assumes no responsibility for any inaccuracies in this document. Juniper
Networks reserves the right to change, modify, transfer, or otherwise revise this publication without notice.
2000384-001-EN Feb 2011 Printed on recycled paper
- Related pdf books
- SDX-300 Installation and Configuration Guide
- Chapter 22
- Security Threat Response Manager Managing Vulnerability Assessment
- APPLICATION NOTE
- QRadar Release Notes - Network Security Solutions ...
- CLI User Guide
- Summary of Attributes in JUNOS XML Tags
- ScreenOS 540 Rev B
- Upgrading to STRM Log Manager 2012 - Network Security ...
- Chapter 15
- Juniper Networks NetScreenHardware Security Client
- BXOS CLI Users Guide and Command Reference
- Topographies for the Simply Connected Campus
- SinglePort T1E1 MiniPhysical Interface Module MiniPIM
- Understanding DAI for Port Security on EXseries Switches
- SSG 20 硬體安裝與組態指南
- Class of Service Configuration Guide - Juniper Networks